PrimeGrid Challenges 2020

[VirtualLarry]

I was going to ask on strategy for GFN 18, 19, and 20. Assuming that they are all supported, by GPUs - will the challenge points that are awarded for completing each of their WUs differ according to their difficulty / run-time length? Would it be advantageous to do only the "shorter" GFN 18 WUs, even though less likely (or perhaps more likely???) to find a prime? Is there any challenge bonus points awarded for finding a prime? (I would assume not, that's just a happy side-effect, but I don't know.)

Also, do the GFN GPU WUs, "prefer" AMD or NVidia? Any significant performance differences? I've been mining on both my AMD RX 5700(XT) fleet, as well as my NVidia GTX 1650 and 1660 ti cards. If one or the other is "better" at PrimeGrid (as the NV cards seem to be better for the purposes of F@H), then I'll leave the other ones mining.

 

[StefanR5R]

When you edit your PrimeGrid project preferences, you will be shown this:
GFN-18: standard credit
GFN-19: standard credit
GFN-20: 10% long job credit bonus

 

[emoga]

Not sure how much help I can provide this round, but if we start losing to the Noobs again I might switch everything over

GFN18 on CPU's
GFN20 on GPU's

I'm guessing that's best option for max points? Has anyone got any performance numbers for GFN 18-20 on cpus?
Edit: Seems like 19-20's on cpus require too much cache for everyday computers. I'll stick with 18's on cpus.

 

[lane42]

Can you bunker for this.

 

[emoga]

Only tasks issued AFTER the start time and returned BEFORE the finish time will be counted.

Link to source

Can't bunker. Probably why a lot people like the PG challenges over FB

 

[StefanR5R]

Seems like 19-20's on cpus require too much cache for everyday computers. I'll stick with 18's on cpus.

Avoid to run too many concurrent jobs:

cpuGFN20, cpuGFN19, cpuGFN18, being single-threaded, are not very viable to run on CPUs in practice. Their data allocations are 16 MB, 8 MB, and 4 MB respectively. Therefore many CPUs will have too small caches to properly support several concurrent tasks or even just one task of these applications, as they become bottle-necked by I/O to main memory.

 

[Ken g6]

Link to source

Can't bunker. Probably why a lot people like the PG challenges over FB

Can't pre-bunker, anyway.

Avoid to run too many concurrent jobs:

In the past, it seemed like adding more tasks, up to the number of real cores, had benefits. It might make sense to leave a full real core for each GPU, or something, though.

 

[StefanR5R]

If you run only CPU tasks on a single machine, then the number of concurrent tasks needs to be a compromise between under-utilization of the vector units and over-utilization of cache and RAM I/O.

It's possible that this compromise is # tasks = # physical cores on many CPUs, if you can live with the then long run times. ... At least CPU temperatures should be pretty low then, because the cores would be mostly waiting for the memory controllers to deliver.

But if you run these CPU tasks on a machine which is also servicing a GPU, then a GPU application which involves a certain amount of copying between RAM and VRAM will suffer quite a bit if the CPU tasks cause a lot of RAM accesses. Most notably Folding@home would suffer; I don't remember how much OCLcudaGFN/ openclatiGFN would be affected.

 

[Ken g6]

We're off! I got off to a "late" start only because I couldn't wake up that early.

In other news, I got some free credit from DigitalOcean. So catch me if you can!

Xeon Golds seem not to have a huge advantage here. I may save some credit for the Tour De Primes in February.

 

[Ken g6]

But if you run these CPU tasks on a machine which is also servicing a GPU, then a GPU application which involves a certain amount of copying between RAM and VRAM will suffer quite a bit if the CPU tasks cause a lot of RAM accesses. Most notably Folding@home would suffer; I don't remember how much OCLcudaGFN/ openclatiGFN would be affected.

I think Folding@home does a lot more transfers between CPU and GPU than your average math application. PrimeGrid GPU apps should be less affected by main RAM speed.

 

[Ken g6]

When you edit your PrimeGrid project preferences, you will be shown this:
GFN-18: standard credit
GFN-19: standard credit
GFN-20: 10% long job credit bonus

There's more to it than that.

GFN-18 WU's are currently averaging 6.5 hours on CPU and 30 minutes on GPU. (GPU is 13 times as fast)
GFN-19 WU's are currently averaging 3.5 days on CPU and 1.75 hours on GPU. (GPU is 48 times as fast)
GFN-20 WU's are currently averaging 25.5 days on CPU and 5 hours on GPU. (GPU is 122.4 times as fast)

The GPU app really does work better on larger numbers. Once you're sure you GPU is stable, you should probably move it to GFN-20.

 

[waffleironhead]

Mini Fleet reporting in, running 18's
3 hours on my Vega 11's.
4 hours on my rx550
6 hours on r7 in my a12-9800

 

[lane42]

GFN18 on cpu

Amd 3800x@4.2, 2.5 Hours
Intel corei3@3.6, 4 Hours
Intel corei9, 9900k@4.6, over 8 hours

 

[Ken g6]

Intel corei9, 9900k@4.6, over 8 hours

You're just doing real cores, right? Set BOINC to do 50% of the threads it sees. If that doesn't help, maybe you can OC your RAM?

My GPU just finished its first GFN-20. It looks like, in terms of points, one GFN-20 ~= 19 GFN-18's. So the points from my GPU vs. my CPUs are pretty close to equal, not counting cloud CPUs. I don't think I want to waste my cloud CPU credit on this challenge too much.

GFN-18s:
i7-6700 (throttled @3.4GHz): 5 hours
Xeon (Haswell) @2.5GHz: ~8.5 hours
Cloud CPUs: 4-8 hours

GTX 1060 3GB: <7 hours for a GFN-20

 

[lane42]

You're just doing real cores, right?

yes. The cpu is a 8/16. I disabled H/T in bios and it's just
running 8 cores. well 7. left 1 open for gpu.

Set BOINC to do 50% of the threads it sees.

yes. That Did very little in the progress bar in B/M.
This is a new cpu for this motherboard. Updated bios.
had a core i5 in it. Guess it still needs work. Same ram as
other computers 3200mhz. Should be fasssssst

 

[Markfw]

yes. The cpu is a 8/16. I disabled H/T in bios and it's just
running 8 cores. well 7. left 1 open for gpu.

yes. That Did very little in the progress bar in B/M.
This is a new cpu for this motherboard. Updated bios.
had a core i5 in it. Guess it still needs work. Same ram as
other computers 3200mhz. Should be fasssssst

Upgrade to Ryzen 5000 series when you can buy one. The old I5 just won't cut it anymore....

 

[lane42]

Mark, this i9 9900k replaced the i5. I think it should be around
2.5 to 3 hours a work unit, not over 8.
waiting on the 5950 and 2 3080ti

 

[Markfw]

Mark, this i9 9900k replaced the i5. I think it should be around
2.5 to 3 hours a work unit, not over 8.
waiting on the 5950 and 2 3080ti

ahhh, yes, I misread. Not a bad CPU, just can't compare to the 5950x I have coming, and you are going to get.

 

[Ken g6]

@Markfw in a PrimeGrid thread?!

This is a new cpu for this motherboard. Updated bios.
had a core i5 in it. Guess it still needs work. Same ram as
other computers 3200mhz. Should be fasssssst

Is the RAM in the proper slots for dual-channel mode?

 

[Kiska]

I wasn't planning on doing the challenge, but today is just 70F/21C, so need to heat my room a bit

 

[lane42]

Is the RAM in the proper slots for dual-channel mode?

yes, tomorrow's another day
run time- cpu time- credit

30,089.44​

29,835.08​

1,723.65​

 

[Kiska]

Oh nevermind, AMD has broken my drivers for compute

 

[Ken g6]

Day 1 stats:

Rank___Credits____Username
76_____641410_____Lane42
160____230159_____biodoc
165____219300_____Ken_g6
187____164399_____emoga
202____144820_____SlangNRox
286____54376______zzuupp
312____37921______geecee
342____27582______waffleironhead

Rank__Credits____Team
20____1984725____BOINCstats
21____1904729____Space Quest
22____1802485____Team Norway
23____1519970____TeAm AnandTech
24____1507190____Ukraine
25____1412694____BOINC@Poland
26____1379545____Iowa State University

Well, that's a better start than last time.

Not sure how much help I can provide this round, but if we start losing to the Noobs again I might switch everything over

The Noobs are about 6 places behind us so far.

 

[StefanR5R]

GFN18 on cpu

= 4 MB hot data per program instance

Amd 3800x@4.2, 2.5 Hours
Intel corei3@3.6, 4 Hours
Intel corei9, 9900k@4.6, over 8 hours

How many concurrent jobs are they running?
8, 2, and 8?

The RAM bandwidth of these processors is at the same magnitude, I guess.

The 3800X has 32 MB last level cache in total = 4 MB per core = borderline enough for one cpuGFN18 instance on each core.
The i3 has got 6 MB LLC in total = 3 MB per core, probably.
The i9-9900K has 16 MB LLC in total = a pitiful¹ 2 MB per core.

_____________
¹) in the context of PrimeGrid's applications

 

[lane42]

Took out the 9900k and put the coffee lake i5 back in, 6 real cores.

The i9-9900K has 16 MB LLC in total = a pitiful¹ 2 MB per core.

Had it running 6 cores, left 2 open for 2 gpu's.
I want to go back to Seti and just GPU'S